Google DeepMind has released AlphaGenome 2, an advanced genomic sequence model engineered to predict the regulatory effects of non-coding DNA mutations at single-cell resolution. Trained on more than 107 million single-cell profiles gathered across 68 human and mouse tissues, the system isolates cell-type-specific regulatory disruptions previously obscured by tissue-level averages. According to Google DeepMind, the model accepts DNA sequences up to one megabase, or one million base pairs, alongside target tissue and cellular state parameters. It quantitatively infers variations in gene expression and chromatin accessibility without requiring additional fine-tuning. Google Chief Executive Sundar Pichai stated that the model clarifies cellular causes of disease by giving scientists tools to anticipate mutation impacts on specific cell types, while DeepMind Vice President of Science Pushmeet Kohli framed it as a next-generation platform for interpreting genetic mechanisms.
Legacy genomic models faced intrinsic limitations because they relied on bulk tissue data containing mixed cell populations, diluting signals that function exclusively in specialized environments like disease-activated or immune cells. Because many disease-associated variants occur in non-coding DNA regions that do not directly specify proteins, detecting transcription disruptions demands precise, single-cell computation. In quantitative evaluations on the TenK10K benchmark encompassing 28 immune cell types, AlphaGenome 2 achieved a 30 percent recall rate in identifying causal expression quantitative trait loci variants, nearly doubling the 16 percent rate recorded by the original AlphaGenome. The model also reached 43 percent accuracy in predicting the directional increase or decrease of gene expression, outperforming the previous leading benchmark baseline of 23 percent by 20 percentage points through coordinated alignment of transcription factor binding and chromatin structural changes.
Despite these performance gains, researchers note limitations: the model recovers only a modest fraction of statistically estimated causal variants within high-precision ranges, and capturing ultra-long-range regulatory interactions beyond one megabase remains an area for improvement. Nevertheless, by allowing research teams to simulate mutation responses in uncharacterized cell states directly from single-cell RNA sequencing data without retraining model weights, AlphaGenome 2 alters in silico discovery pipelines. Investigators can computationally reconstruct pathological mechanisms, such as gene overactivation in gut immune cells linked to intestinal disease, shortening traditional wet-lab experimental cycles. DeepMind is offering free AlphaGenome API access to non-commercial academic researchers while launching private preview access for commercial enterprises and pharmaceutical developers.